๐ NLP Understanding¶
๐ฌ EMNLP2026 ยท 49 paper notes
๐ Same area in other venues: ๐ฌ ICLR2026 (2) ยท ๐ฌ ACL2026 (34) ยท ๐งช ICML2026 (2) ยท ๐ค AAAI2026 (1) ยท ๐ง NeurIPS2025 (3) ยท ๐น ICCV2025 (1)
๐ฅ Top topics: LLM ร13 ยท Question Answering ร5 ยท Alignment/RLHF ร2
- A Survey on Representing Linguistic Style: Challenges and Opportunities
-
This survey synthesizes linguistic theories and NLP representation learning to establish a unifying, content-independent definition of style, reviews predefined and neural style embedding architectures, and empirically shows that dedicated style embeddings achieve superior accuracy over frontier LLMs at thousands of times lower compute.
- Alice: A Large-Scale German Benchmark for Rubric-Based Multi-Dimensional Automatic Short Answer Scoring
-
Addressing the limitations of conventional automatic short answer scoring that rely on reference answer overlap and single global scores, this paper introduces ALICE, a 16.5K German student response benchmark covering three distinct dimensions (learning performance, knowledge elements, and scientific skills), and formulates scoring as cross-encoder rubric retrieval to dramatically improve generalisation across unseen questions.
- Are Non-English Papers Reviewed Fairly? Language-of-Study Bias in NLP Peer Reviews
-
This paper provides the first systematic characterization of language-of-study (LoS) bias in NLP peer review, presenting the expert-annotated LOBSTER benchmark alongside an LLM detection pipeline (87.37 Macro F1), and empirically demonstrates across 15,645 reviews that non-English papers face ~40x higher bias rates than English-only papers, dominated by unwarranted demands for cross-lingual generalizability.
- BeyondAtoms: Benchmarking and Verifying Relational Factuality Beyond Atomic Facts
-
Addressing the vulnerability where conventional fact verifiers and atomic decomposition blindly trust entity/event overlaps while ignoring logical topologies, this paper introduces BEYONDATOMS, the first operation-centric relational factuality benchmark, and CARVE, a training-free verification framework based on claim-anchored extraction and bidirectional structural alignment.
- CausalTAD: Injecting Causal Knowledge into Large Language Models for Tabular Anomaly Detection
-
Addressing the flaw where random column ordering disrupts feature causal dependencies in LLM-based tabular anomaly detection, CausalTAD proposes high-level latent factor causal discovery, column-level causal preference projection, linear ordering problem (LOP) relaxation, and factor-count reweighting, achieving state-of-the-art performance across 30+ benchmarks.
- charting empirical patterns for llm fine tuning in scientific multi discipline l
-
A systematic empirical study of LLM joint fine-tuning across five naturally imbalanced scientific disciplines (math, chemistry, biology, medicine, and geography) distills four governing patterns and shows that an asymmetric shared-A LoRA-MoE rivals full fine-tuning while training only 0.902% of the parameters.
- CodeTD: Topology of Attention Detects Hallucinations in Code LLMs
-
CodeTD models the attention maps of Code LLMs as weighted graphs, leverages persistent homology and Manifold Topology Divergence (MTD) to extract multi-scale 0-dimensional connectivity and 1-dimensional detour cycle features between prompt and generation tokens, enabling highly accurate, execution-free code hallucination detection and best-of-N reranking.
- Deep and shallow biases in language models
-
This paper introduces the bias depth score \(\pi = \text{DR} \times \text{FR}\), decoupling repeated LLM preferences in open-ended choice tasks into persistent Deep biases inherited from pretraining and fragile Shallow biases driven by surface prompt wording across 4,442 grounded prompt families.
- DiscoTrace: Representing and Comparing Answering Strategies of Humans and LLMs in Information-Seeking Question Answering
-
DiscoTrace represents long-form answers as structured sequences of question-related discourse acts paired with explicit question interpretations, revealing that human communities exhibit diverse rhetorical answering policies while large language models remain rhetorically homogeneous and systematically over-answer low-probability interpretations.
- DiscoTrack: A Multilingual LLM Benchmark for Discourse Tracking
-
DiscoTrack establishes a standardized question-answering benchmark spanning 12 typologically diverse languages and 11 subtasks across four discourse tiers (salience, entity tracking, discourse relations, bridging inference), demonstrating that even frontier LLMs struggle with implicit document-level discourse reasoning.
- Does Finetuning with Scientific Data Increase Hallucinations? A Multi-domain Factuality Evaluation of LLMs
-
Introduces the SCIFACTCHECK benchmark spanning 2,500 concepts across five scientific domains and a fine-grained evaluation framework targeting unverifiability, overclaiming, and attribution, empirically demonstrating via controlled minimal-pairing that scientific fine-tuning degrades factual reliability and exacerbates a dissociation where models exhibit higher linguistic certainty despite lower internal parametric confidence.
- Encoder-only and decoder-only models exhibit distinct patterns of linguistic bias
-
Using the strictly paired Ettin model suite alongside standard open-source language models, this paper investigates how training objectives shape linguistic bias in semantic role assignment and structural voice choice, revealing that while all models exhibit a strong animacy-first heuristic, bidirectional encoders generalize verb argument structures more flexibly and prefer passives compared to frequency-bound autoregressive decoders.
- Evaluating Communicative Belief Updates in Large Language Models via Implicature Recognition and Cancellation
-
This paper introduces IMPLICATUREX, the first expert-annotated and crowdsourced benchmark formalizing communicative belief updates through implicature recognition and cancellation, revealing that leading LLMs lag substantially behind humans in naturally-occurring dialogues and suffer from severe belief inertia and prior memorization.
- Exposรญa: Teaching and Assessment of Academic Writing Skills for Research Project Proposals and Peer Feedback
-
This paper introduces Exposรญa, the first publicly available benchmark dataset collected from an authentic university computer science course that unites student research proposal drafts, span-anchored inline comments, free-text peer and instructor reviews, final revisions, and human scores across fine-grained pedagogical rubrics, establishing comprehensive zero-shot LLM baselines under binned JensenโShannon divergence.
- From Prompts to Trees: Effective LLM-Guided Tree Generation for Few-Shot Tabular Classification
-
This paper introduces LLMT, a three-stage distillation framework that extracts prior knowledge from Large Language Models into compact, interpretable decision trees by generating discrete rules, statistically assembling them with few-shot Gini criteria, and performing leaf-level LLM verification, achieving high classification accuracy with zero LLM inference cost.
- From Text to Structure: The PACKaGE Benchmark for Bidirectional Causal Reasoning in Paragraphs and Graphs
-
This paper introduces PACKaGE, the first bidirectional causal reasoning benchmark pairing paragraph-level natural language narratives with explicit Directed Acyclic Graphs (DAGs) across 4,566 curated instances, uncovering a profound capability dissociation between causal extraction and causal generation in modern LLMs.
- global consistency repair for natural language claim sets
-
This paper formalizes global consistency repair for natural-language claim sets and introduces QXR, an adaptive diagnosis-and-repair framework that combines black-box LLM consistency judgments with recursive minimal unsatisfiable subset localization and greedy hitting-set deletion, substantially mitigating over-pruning while preserving maximum coherent information.
- Granuscore: A Reference-Free Measure of Granularity for Text Analysis and Question Answering
-
This paper introduces Granuscore, a reference-free measure of semantic granularity that combines the radial depth of hyperbolic Poincarรฉ embeddings with anchor manifold relations via gradient-boosted decision trees and percentile calibration, uncovering how granularity disparities drive question-answering difficulty and model error patterns.
- h table hierarchical hypergraph enhanced large language models for complex table
-
To resolve the structural loss and dependency collapse caused by serialization and flat hypergraph modeling on complex multi-level nested table headers, HยฒTable formalizes complex tables as hierarchical nested hypergraphs, coupled with a four-stage hierarchy-aware hypergraph encoder and a learnable query-based structural alignment module, achieving marked improvements in TableQA on deep hierarchical tables while updating only ~1% of parameters.
- HALT: Hallucination Assessment via Log-probs as Time series
-
HALT formulates an LLM's top-20 log-probability sequence from a single generation as a multivariate time series, combines it with 5 engineered local calibration features, and uses a compact 5M-parameter bidirectional GRU with Top-q pooling to deliver highly accurate, low-cost grey-box hallucination assessment.
- How LLMs Follow Instructions: Skillful Coordination, Not a Universal Mechanism
-
Through diagnostic probing, cross-task transfer, INLP causal ablation, and longitudinal temporal analysis across nine diverse instruction tasks in three open-source LLMs, this work demonstrates that instruction following is driven by the dynamic coordination of localized linguistic skills rather than a universal, task-independent constraint satisfaction mechanism.
- Human Label Variation as Stable Signal: Learning Annotator-Specific Explanation Behavior via Cross-Annotator Preference Optimization
-
Addressing the limitation of conventional human label variation (HLV) research that models only categorical label distributions, this paper introduces Cross-Annotator Preference Optimization (CAPO), leveraging pairwise contrasts between different valid annotators on identical inputs as preference supervision to teach LLMs annotator-specific reasoning and justification habits.
- HyperProve: Answer-Guided Hypergraph Expansion for Multi-Hop Question Answering
-
Addressing evidence drift and context collapse in multi-hop QA, HyperProve organizes corpus evidence into a synonym-aware hypergraph of atomic facts, carries validated intermediate answers and supporting hyperedges as structural retrieval state, and performs query-biased local random-walk expansion to construct faithful multi-hop evidence chains.
- Investigating and Enhancing Concept Induction in LLMs: A Prototype-based Property-augmented Approach
-
Addressing the pervasive problem where large language models systematically produce overly general concepts during semantic abstraction, this paper establishes a familiarity-controlled evaluation benchmark via QA-based knowledge probing and proposes PPACI, a prototype-based property-augmented concept induction approach that fuses intensional definitions with extensional member properties and provides dual-level candidate anchors, reducing overly general predictions by 17โ28 percentage points while improving induction accuracy.
- JustifIT-QA: Evidence-Grounded Legal Question Answering in Italian
-
Introduces JustifIT-QA, the first Italian statutory QA dataset featuring expert-authored answers and fine-grained legal evidence spans, alongside an end-to-end framework combining the domain-adapted dense retriever DARv2, a judge-driven conversational refinement loop, and token containment-based evidence extraction.
- LAMPEAE: LLM-Augmented Manifold Probing for Adaptive Event Argument Extraction
-
Addressing the geometric rigidity of static role anchors in prompt-based event argument extraction, LAMPEAE leverages a frozen LLM to distill high-order reasoning priors and dynamically warps static anchors along Riemannian manifold trajectories via a low-rank hypernetwork, achieving a new state-of-the-art Arg-C F1 of 54.5% on the RAMS benchmark.
- Large Language Models Systematically Favor Popular Options: Evidence, Mechanisms, and Mitigation
-
The paper reveals that LLMs systematically favor popular but incorrect options in multiple-choice questions (popularity bias), introduces PopMCQ across six controlled strategies, and proposes PopDebiasโa lightweight, inference-time prior-removal method that boosts accuracy by up to 31.9 percentage points without retraining.
- Learning the Cue or Learning the Word? Analyzing Generalization in Metaphor Detection for Verbs
-
Using a controlled lexical hold-out benchmark and geometric embedding probes on English verbs, this paper proves that transformer encoders generalize in metaphor detection predominantly by learning transferable contextual cues, whereas target verb identity serves merely as an additive memory boost.
- LLM Feedback for Topic Model with Preference Optimization and Contrastive Learning
-
LLMFeed is an efficient neural topic modeling framework that incorporates semantic feedback from an LLM through a single offline interaction phase, using Direct Preference Optimization (DPO) to refine topic words and topic-guided contrastive learning to align document-topic distributions, achieving substantial gains in both coherence and clustering quality at minimal computational cost.
- llm summaries systematically distort topic distributions
-
A systematic benchmark reveals that large language models systematically and stubbornly distort source topic distributions in multi-document summarization, persistently amplifying topics aligned with their pre-trained importance priorsโeven when entirely absent from input reviewsโand resistant to mitigation via scaling or prompt engineering.
- lsc pattern interpretable lexical semantic change analysis via diachronic sense
-
LSC-Pattern proposes an interpretable framework for lexical semantic change analysis that couples dictionary-grounded diachronic sense distributions with operationalized semantic dimensions, leveraging LLM-based word sense disambiguation to ground usages into lexicographic senses and decomposing macro-trajectories across valence, semantic scope, and metaphoricity into concrete sense-frequency shifts.
- Measuring Data Diversity with Embeddings: A Taxonomy and Benchmark for Diversity Measures
-
This paper unifies 21 embedding-based diversity measures into four distinct mathematical families and introduces EmbDivBench, the first controlled benchmark evaluating measures across variety, balance, and disparity on both simulated Gaussian mixtures and natural Wikipedia text, demonstrating that the ubiquitously used Mean Pairwise Distance underperforms while Graph Entropy provides the best overall performance.
- MoF: Preference-Aware Mixture Modeling for Black-Box LLM Personalization
-
Addressing the parameter explosion and cold-start adaptation bottlenecks of user-specific scoring heads in black-box LLM personalization, MoF models user preferences as adaptive sparse compositions over shared latent facet heads guided by a history-conditioned sparse autoencoder (SAE) router and cluster-based sampling (CBS), achieving constant \(O(1)\) personalization parameters and strong generalization to unseen users without fine-tuning.
- Not All or None: Dynamic Construction of Target-aware Memory Graph for Conversational Stance Detection
-
TamGraph resolves the dilemma between noisy full history and context-starved single-turn inputs in conversational stance detection by dynamically extracting target-related propositions and building a stance memory graph via entropy-guided stepwise backtracking, substantially outperforming single-turn, multi-turn, and CoT baselines on both MT-CSD and ZS-CSD benchmarks.
- On the Optimality of Kinship Naming: an Information-theoretic Approach
-
This paper adapts the Information Bottleneck principle to discrete semantic naming by defining information loss via multiclass cross-entropy and deriving a need-agnostic theoretical optimal frontier \(L = -\tilde{C}\), proving that kinship naming in four natural languages achieves theoretical optimality under Bayesian listeners and demonstrating that neural multi-agent communication autonomously evolves toward this Pareto optimal boundary.
- Ontology-Driven Structural Regularization for Document-Level Relation Extraction
-
Targeting severe structural logical contradictions in distantly supervised datasets for document-level relation extraction, this paper proposes an ontology-driven structural regularization framework grounded in Web Ontology Language (OWL) axioms that filters ill-formed entities and completes missing inverse relations prior to training, significantly suppressing prediction contradictions and substantially boosting extraction F1 without architectural modifications.
- OVLAFND: Open-Vocabulary Logic-Aware Fake News Detection
-
This paper proposes OVLAFND, which constructs counterfactual hard negatives via multimodal LLMs to learn fine-grained logic consistency in a dedicated subspace, and replaces rigid domain labels with an open-vocabulary topic manifold aligned to semantic anchors, substantially boosting fake news detection accuracy and zero-shot cross-domain generalization.
- Promptology: A Large-Scale Systematic Analysis of Prompt Elements and Contextual Factors in Argument Generation
-
Conducting the first fully factorial analysis of political argument generation across three languages with over 375k machine arguments and human politician baselines, this work demonstrates that LLMs remain highly separable from human authors in style and semantics (>90% balanced accuracy), where sociodemographic persona prompting is the only element that moderately bridges the gap, while model architecture, language, and topic dominate generation variance and output stability.
- Quizzing the Translation: A Prover-Grounded Evaluation Metric for NL->FOL
-
Existing surface-level and graph-based metrics fail to track logical severity and often assign higher scores to more broken formulas; SIV quizzes candidate translations against prover-grounded positive and contrastive probes derived from the reference, delivering calibrated graded scores and diagnostic error traces.
- Relational Linearity is a Predictor of Hallucinations
-
This paper introduces and empirically validates the relational linearity hypothesis, demonstrating that the representational accessibility of relation-level affine maps (\(\Delta\cos\)) is a powerful predictor of whether language models hallucinate plausible values instead of refusing when queried on unknown entities (\(r \in [0.58, 0.84]\) across multiple model families).
- Rubric Spans are Label Representations: Joint LLM Encoding for Short Answer Scoring
-
To break the trade-off between the rigidity of fixed classification heads and the steep compute overhead of pairwise retrieval or autoregressive generation in short answer scoring, RUSPAN encodes question contexts, answers, and all candidate rubric spans in a single decoder-only LM pass to perform listwise scoring, while Rubric-Independent Masking (RIM) and position reindexing eliminate inter-rubric attention dependencies to achieve robust cross-benchmark zero-shot transfer.
- Structure-Aware Data Augmentation for Argument Mining
-
This paper proposes a structure-aware data augmentation framework for argument mining that samples diverse pseudo-argumentative tree topologies using a PCFG learned from training data, linearizes them into TANL templates, and guides an LLM to generate text under structure-aware demonstration selection and automated diagnostic feedback, significantly boosting component and relation extraction performance.
- SustainableQA: A Comprehensive Question Answering Dataset for Corporate Sustainability and EU Taxonomy Reporting
-
To tackle the severe information fragmentation, complex multi-page tables, and strict regulatory terminology in corporate sustainability reports, SustainableQA introduces a comprehensive benchmark of 195,287 QA pairs generated via a hybrid span extraction and table-to-paragraph pipeline, safeguarded by deterministic numeric checks and dual-metric adaptive refinement, enabling a fine-tuned 8B model to outperform 70B and proprietary frontier baselines.
- Syntactic Prediction through Reinforcement Learning
-
Addressing the "supervised plateau" where supervised fine-tuning (SFT) fails to enable large language models to surpass specialized neural syntactic parsers, this paper proposes Syntax-RLโa two-stage post-training framework that establishes unified syntactic sequence generation via cold-start SFT on 80% data and directly optimizes structure-level objectives (bracket F1 and token accuracy) via GRPO on a disjoint 20% subset, setting new state-of-the-art records on constituency parsing and CCG supertagging.
- thinking in different spaces domain specific latent geometry survives cross arch
-
By learning an affine projection between independently trained heterogeneous LLMs and intervening in the student residual stream during inference, this work demonstrates that cross-architecture latent geometry can be linearly aligned for zero-weight-update behavioral correction, while proving that factual recall and mathematical reasoning occupy geometrically orthogonal domain subspaces.
- Usage-Modulated Sentiment Representations in Large Language Models
-
Addressing whether sentiment inside large language models is governed purely by a single linear direction, this paper introduces a counterfactually controlled paired dataset holding event skeletons fixed across tone and audience variations, showing that orthogonal residual vectors retain a compact, reproducible usage subspace that causally steers generation tone while preserving sentiment polarity.
- What Can Reviews Reliably Tell Us About Books? A Per-Attribute Audit and Tiered Annotated Dataset Across 1.7M Titles
-
Auditing eight scalable review-derived feature methods against a controlled human reading benchmark across 391 narrative attributes reveals that most apparent predictive accuracy is popularity-confounded, with only 12 attributes establishing true content-driven recovery governed by lexical anchorability.
- When Does Auxiliary Classification Help Regression in Practice? A Case Study in Arabic Dialectness Modeling
-
This paper investigates why naive auxiliary classification degrades continuous regression under heavily skewed target distributions and demonstrates that combining 7-bin equal-width discretization with Focal Loss turns harmful negative transfer into a strong regularizer, achieving a state-of-the-art RMSE of 0.1275 on the NADI 2024 Arabic dialectness benchmark (over 9% improvement over prior systems) and transferring directly to English toxicity regression with zero inference overhead.
- when should models change their minds contextual belief management in large lang
-
This paper introduces the Contextual Belief Management (CBM) framework and BeliefTrack, a closed-world benchmark spanning Rule Discovery and Circuit Diagnosis, demonstrating that frontier LLMs suffer severe failures in belief preservation, updating, and isolation, while verifier-guided reinforcement learning with dense Jaccard belief-state rewards and representation steering significantly mitigate these failures.